On a 64-core AMD EPYC 9575F, Reamer Research replays about 1.7 million bars a second in one backtest, and Reamer Server takes an order from a strategy to an accepted event in a median of 11.3 µs, with about 200,000 to 220,000 orders a second across many strategies.
Every figure below is from the kits' own benchmark documents or the whitepaper. Each one measures something specific, so read the conditions with the number.
Reamer Research
- About 1.7 million bars a second on the EPYC, with a strategy that places no orders, called through the C interface. Across 2,000 runs, the slowest was within 4% of the fastest.
- With real trading, a 20-bar breakout strategy making 47,728 trades over 884,130 bars took about 1.9 seconds from C or C++ and about 65 seconds from Python. That was measured on a 2017 laptop processor, so a server will be faster.
- Python costs about 33 times more per bar than C++, because each bar crosses into Python.
- Scale: on the EPYC, 50 million daily bars across 10,000 instruments over 20 years, with 990,000 orders carrying stop-loss and take-profit brackets, ran in 29.1 seconds using about 1 GB of extra memory. The bars were generated for the test.
Each backtest runs on one thread. For a parameter sweep, run several backtests at once.
Reamer Server
- 11.3 µs median and 12.4 µs at the 99th percentile, from a strategy sending an order over the local socket to its accepted event, with one strategy, an accept-all gate and a stub connector.
- About 200,000 to 220,000 orders a second in total, with roughly 32 to 67 strategies on the one instance.
- Latency stays steady up to about one strategy per physical core. Past that, the slowest orders get slower first.
These figures leave out your gate's own logic, the network and the broker. A broker round trip typically takes 1 to 10 milliseconds or more, so it outweighs the server's share. Neither product is built for high-frequency trading.
Measure it yourself
The Research kit ships research-bench and the Server kit ships server-bench. Each runs the benchmark above on your own machine in one command, during the 30-day trial. The method and full results from the EPYC are in the whitepaper.